期刊
SCIENTIFIC REPORTS
卷 12, 期 1, 页码 -出版社
NATURE PORTFOLIO
DOI: 10.1038/s41598-022-15855-7
关键词
-
资金
- National Science Foundation of China [61703355]
- Natural Science Foundation Project of Zhangzhou 2019 [ZZ2019J34]
- Program for Young Excellent Talents in University of Fujian Province [201847]
Surface defect detection is a vital process in industrial production and a significant research direction in computer vision. This paper proposes a new unsupervised small sample defect detection model that improves detection accuracy and defect profile extraction by enhancing the feature extraction capability of the generator and introducing an SSIM-based defect segmentation method.
Surface defect detection is a vital process in industrial production and a significant research direction in computer vision. Although today's deep learning defect detection methods based on computer vision can achieve high detection accuracy, they are mainly based on supervised learning. They require many defect samples to train the model, which is not compatible with the current situation that industrial defect sample is difficult to obtain and costly to label. So we propose a new unsupervised small sample defect detection model-ISU-GAN, which is based on the CycleGAN architecture. A skip connection, SE module, and Involution module are added to the Generator, enabling the feature extraction capability of the model to be significantly improved. Moreover, we propose an SSIM-based defect segmentation method that applies to GAN-based defect detection and can accurately extract defect contours without the need for redundant noise reduction post-processing. Experiments on the DAGM2007 dataset show that the unsupervised ISU-GAN can achieve higher detection accuracy and finer defect profiles with less than 1/3 of the unlabelled training data than the supervised model with the full training set. Relative to the supervised segmentation models UNet and ResUNet++ with more training samples, our model improves the detection accuracy by 2.84% and 0.41% respectively and the F1 score by 0.025 and 0.0012 respectively. In addition, the predicted profile obtained using our method is closer to the real profile than other models used for comparison.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据